Steroid and Non-Steroid Bodybuilder Classification Using Deep-Learning

This paper aims to implement an image classification system to detect Steroid and Non-steroid bodybuilders with RGB images using deep learning techniques. The purpose of creating a steroid use detection Deep Learning system is to provide a reliable dope-testing alternative for sports organization committees and athletes without the need for expensive time-consuming lab tests that can easily be bribed out of. Moreover, this allows a method to avert the failure of detecting steroid use in bodybuilders, as its use leads to unfair competitive advantages and life-threatening health risks. Particularly, cardiovascular diseases like high blood pressure and stroke, infections like hepatitis, and HIV, and psychiatric effects like mania, and addiction. While some side effects may be semi-permanent and reversible once the drug use is stopped some leave irreparable life-long damage. Therefore, we have created a CNN-based Deep-learning steroid detection system that doesn't rely on people but on data and proof. Firstly, a non-steroid and steroid bodybuilder image dataset of 4,231 RGB images was collected and cleaned. Next, we trained multiple pre-trained CNN models using this dataset. Lastly, we have reached reliable results with our pre-trained models, namely, our highest accuracy being 98.00% with InceptionV3 of epoch 30, and our lowest accuracy is 94.00% with Resnet50 of epoch 50, all other model results fall into this range. Finally, this is how our project provides an accurate Deep-learning Image classification system capable of steroid and non-steroid bodybuilder detection.